A Better Brainstorming System: Diagnose Convergence Without Mistaking Consensus for Quality

The complete answer to why AI brainstorming can converge too early is now visible: a model’s shared priors can make common directions easy to sample, sequential context can anchor each session to its own early output, and interaction can synchronize agents around an authoritative or already popular frame. These forces can operate at different stages and reinforce one another before anyone selects a proposal. The practical judgment is therefore not whether an AI system agrees with itself, but whether it preserved enough independent, interpretable alternatives for agreement to mean something.

The complete causal answer

The first distinction is between fluency and range. A language model can write a coherent answer because it is good at continuing a context. That same strength can make a first premise unusually easy to elaborate. A polished list may contain several descriptions of one route through the problem rather than several routes. The issue is not a lack of words. It is a shortage of independent information before selection.

At the level of one session, fixation creates that shortage through sequence. An early idea remains in view, related continuations become salient, and later ideas inherit the first idea’s structure. Alignment toward familiar or conventional responses can reinforce the trajectory. The observable result is slower accumulation of new categories or combinations, not proof that the model has a human psychological state.

At the level of independent sessions, knowledge aggregation creates a different shortage. Human contributors are individually limited, but their experiences and associations are partitioned across different minds. They may begin from different regions and follow different paths. LLMs aggregate broad knowledge and preference signals into a shared system, so many sessions can access a large corpus while still sampling repeatedly from similar central regions. A model can know about unusual possibilities without making those possibilities equally likely to appear in every independent run.

At the level of a multi-agent group, structural coupling adds a third pressure. Once agents see a leader’s framing, a peer’s proposal, or a common discussion history, their next contributions are conditioned on information that was not present for the others at the beginning. Authority can turn an initial frame into a gravitational center. More agents can add less new information as their outputs overlap. More discussion can add detail around a stable center while reducing the range of centers the group considers.

The forces are cumulative, but they are not identical. A system may have a broad first-pass pool and then collapse under a leader-led discussion. It may have narrow first ideas but relatively independent later elaboration. It may anchor each persona’s session while personas keep the population diverse. Or it may fail at every level. The phrase “the AI keeps giving us the same idea” is too vague to determine which of these cases is occurring.

The two controlled research programs support this layered answer from different angles. In the fitness-product studies, default LLM sessions were not dramatically more fixated within ten sequential ideas than human sessions, but their first ideas and pooled outputs were less diverse. Human seeds did not improve the LLMs’ downstream diversity, suggesting that a new entry point did not recreate a different associative trajectory. Ordinary heterogeneous personas improved between-session variation but increased within-session fixation, and staged CoT revision reduced that fixation in LLMs.

In the multi-agent study, the same relationship appears as a systems property. The more diverse horizontal interaction structure outperformed authority-weighted structures on semantic diversity, even though quality differences were modest. Increasing group size increased absolute diversity but reduced the amount of new diversity per agent. Debate rounds could become more dispersed within a session while the group’s center stabilized. These results do not say that an LLM or a leader is intrinsically defective. They say that coordination changes the information the system can still discover.